**Mamba-UNet: A Novel Architecture for Medical Image Segmentation** In the field of medical image analysis, Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) have achieved significant advancements. While CNNs excel in capturing local features, ViTs are effective in understanding global context through self-attention mechanisms. However, both architectures struggle with modeling long-range dependencies in medical images, which is crucial for precise segmentation. Inspired by the Mamba architecture, known for its efficiency in handling long sequences and global contextual information, the authors propose Mamba-UNet, a novel architecture that integrates the U-Net with Mamba's capabilities. Mamba-UNet employs a pure Visual Mamba (VMamba)-based encoder-decoder structure, incorporating skip connections to preserve spatial information across different scales. This design facilitates comprehensive feature learning, capturing intricate details and broader semantic contexts in medical images. A novel integration mechanism within the VMamba blocks ensures seamless connectivity and information flow between the encoder and decoder paths, enhancing segmentation performance. Experiments on the ACDC MRI Cardiac segmentation dataset and the Synapse CT Abdomen segmentation dataset show that Mamba-UNet outperforms several types of UNet under the same hyper-parameter settings. The source code and baseline implementations are available on GitHub. The paper discusses the evolution of U-Net with various network blocks and the integration of ViT and State Space Models (SSMs) to improve long-range dependency modeling. The Mamba-UNet architecture is detailed, including the VSS block, encoder, decoder, and skip connections. The implementation details, evaluation metrics, and qualitative and quantitative results are provided, demonstrating Mamba-UNet's superior performance compared to baseline methods. future work aims to explore more medical image segmentation tasks from different modalities and targets, extend Mamba-UNet to 3D medical images, and investigate semi/weakly-supervised learning to further enhance its capabilities.**Mamba-UNet: A Novel Architecture for Medical Image Segmentation** In the field of medical image analysis, Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) have achieved significant advancements. While CNNs excel in capturing local features, ViTs are effective in understanding global context through self-attention mechanisms. However, both architectures struggle with modeling long-range dependencies in medical images, which is crucial for precise segmentation. Inspired by the Mamba architecture, known for its efficiency in handling long sequences and global contextual information, the authors propose Mamba-UNet, a novel architecture that integrates the U-Net with Mamba's capabilities. Mamba-UNet employs a pure Visual Mamba (VMamba)-based encoder-decoder structure, incorporating skip connections to preserve spatial information across different scales. This design facilitates comprehensive feature learning, capturing intricate details and broader semantic contexts in medical images. A novel integration mechanism within the VMamba blocks ensures seamless connectivity and information flow between the encoder and decoder paths, enhancing segmentation performance. Experiments on the ACDC MRI Cardiac segmentation dataset and the Synapse CT Abdomen segmentation dataset show that Mamba-UNet outperforms several types of UNet under the same hyper-parameter settings. The source code and baseline implementations are available on GitHub. The paper discusses the evolution of U-Net with various network blocks and the integration of ViT and State Space Models (SSMs) to improve long-range dependency modeling. The Mamba-UNet architecture is detailed, including the VSS block, encoder, decoder, and skip connections. The implementation details, evaluation metrics, and qualitative and quantitative results are provided, demonstrating Mamba-UNet's superior performance compared to baseline methods. future work aims to explore more medical image segmentation tasks from different modalities and targets, extend Mamba-UNet to 3D medical images, and investigate semi/weakly-supervised learning to further enhance its capabilities.